`rl_global_batch` becomes zero with large `_world_size` during 32B model training
Open
- Dominant language
- Python
- Stars
- 190
- Forks
- 6
- PR merge metrics
- No merged PRs in 30d
Description
https://github.com/InternLM/OREAL/blob/133434bc45c7da4725b0d3fa71ae1ba705bb8c14/train_oreal.py#L598-L613
When training large models (especially 32B parameter models) with distributed processing, there's a potential issue where `rl_global_batch` can become zero if `_world_size` is large. This causes a ZeroDivisionError in the code. Is there any reasonable method to fix this problem?
Contributor guide
No contributing guide indexed for this repository
Assessment
This issue has not been assessed yet.